Learning in Pursuit-Evasion Differential Games Using Reinforcement Fuzzy Learning.

摘要

In this thesis, Q-learning fuzzy inference system is applied to pursuit-evasion differential games. The suggested technique allows both the evader and the pursuer to learn their optimal strategies simultaneously. Reinforcement learning is used to autonomously tune the input parameters and the fuzzy rules of a fuzzy controller for both the evader and the pursuer. We focus more on the behaviours and the strategies of the trained evader. The evader is trained to find its optimal strategy from the received rewards during the game. The homicidal chauffeur game and the game of two cars are used as examples of the method. The simulation results of the suggested technique demonstrate that the trained evader is able to learn its optimal strategies. Furthermore, the learning speed is investigated when using Q-learning with eligibility traces in pursuit-evasion differential games.